Power transmission line short circuit identification method and equipment
Through the recognition method based on sound characteristics, short-circuit discharge of the transmission line is identified using the Mel spectrogram and Mel cepspectral coefficient characteristics, the problem of insufficient accuracy and reliability in the prior art is solved, and the problem of under various environmental conditions is realized. Stable testing reduces costs.
Patent Information
- Application Number
- CN202411085545.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-11-29
AI Technical Summary
The existing short-circuit discharge fault detection methods for power transmission lines are insufficiently accurate and reliable, and are greatly affected by noise and environmental factors, and the equipment costs are high, making it difficult to detect stably under various environmental conditions.
Using a sound feature-based recognition method, by collecting sound information from the transmission line, extracting the Mel spectrogram features and Mel cepspectral coefficient features, constructing fusion voiceprint features, and inputting a pre-trained short-circuit recognition neural network model for identification and judgment Whether it is a short-circuit discharge.
It improves the accuracy and reliability of short-circuit discharge fault detection, reduces detection costs, and can conduct stable detection under various environmental conditions, reduces misjudgment and inaccurate detection, and forms a more comprehensive detection system.
Smart Images

Figure CN119044822A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computer technology, and more particularly to a method and device for identifying a short circuit in a power transmission line. Background Art
[0002] With the continuous development of power systems, the safe operation of transmission lines has become crucial. However, short-circuit discharge may occur during the operation of transmission lines, which poses a serious threat to the normal operation of power systems. Short-circuit discharge failures not only lead to power supply interruptions, but may also cause safety accidents such as fires, causing huge economic losses and social impacts.
[0003] In order to detect and eliminate short-circuit discharge faults in transmission lines in a timely manner, researchers have conducted a lot of research work. At present, some methods have been applied to the detection of short-circuit discharge faults in transmission lines. Among them, using the instantaneous value of current or the true effective value method as the basis for short-circuit judgment is a common method. This method determines whether a short-circuit discharge fault occurs by monitoring the current changes in the transmission line. When the instantaneous value or true effective value of the current exceeds the set threshold, the system will issue an alarm signal.
[0004] In addition, detecting the change of voltage and phase angle during short circuit is also an effective method. When a short circuit discharge fault occurs in the transmission line, the voltage and phase angle will change significantly. By monitoring and analyzing these changes, it can be determined whether there is a short circuit discharge fault.
[0005] In recent years, detection methods based on machine vision have gradually attracted attention. This method uses cameras and other equipment to collect image information of power transmission lines, and uses image processing and analysis technology to detect whether there are abnormal sound, light, heat and other phenomena, so as to determine whether a short-circuit discharge fault has occurred.
[0006] Although these methods can detect short-circuit discharge faults in transmission lines to a certain extent, they still have some limitations. For example, the instantaneous value of current or true effective value method may be affected by noise and interference, leading to misjudgment; the voltage and phase angle change detection method requires high equipment accuracy and is relatively expensive; the machine vision detection method is greatly affected by environmental factors, such as light, fog, etc., which may affect the accuracy of detection. Therefore, a more accurate, reliable and economical transmission line short-circuit discharge fault detection method is needed to improve the safety and reliability of the power system. Summary of the invention
[0007] One or more embodiments of the present specification provide a method and device for identifying a short circuit in a power transmission line, which are used to solve the technical problems raised by the background technology.
[0008] One or more embodiments of this specification adopt the following technical solutions:
[0009] One or more embodiments of this specification provide a method for identifying a short circuit in a transmission line, the method comprising:
[0010] Collect current sound information within the preset range of the transmission line;
[0011] Extracting current Mel-spectrogram features and current Mel-cepstral coefficient features of the current sound information;
[0012] Constructing a current fused voiceprint feature according to the current Mel-frequency spectrum feature and the current Mel-frequency cepstral coefficient feature;
[0013] The current fused voiceprint feature is input into a pre-trained short-circuit recognition neural network model, and a recognition result is output, so as to determine whether the current sound information is generated by short-circuit discharge according to the recognition result.
[0014] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0015] Improved accuracy: By collecting the current sound information within the preset range of the transmission line, extracting its Mel spectrum features and Mel cepstral coefficient features, constructing the current fused voiceprint features, and then inputting them into the pre-trained short-circuit recognition neural network model for recognition. This recognition method based on sound features can more accurately determine whether the sound is caused by short-circuit discharge, avoiding the noise and interference that the instantaneous value of current or the true effective value method may be affected by, as well as the high requirements for equipment accuracy of the voltage and phase angle change detection method, thereby improving the accuracy of short-circuit discharge fault detection.
[0016] Enhanced reliability: This method uses sound information to identify short circuits, which is not affected by environmental factors such as light and fog. Compared with machine vision detection methods, it has higher reliability. It can stably detect under various environmental conditions, reducing misjudgments and inaccurate detection caused by environmental factors.
[0017] Economic advantage: Compared with the voltage and phase angle change detection method, this method may not require high-precision expensive equipment, reducing the detection cost. At the same time, detection through sound information is relatively simple, which also helps to reduce maintenance and operating costs.
[0018] Good real-time performance: It can collect and analyze the sound information of the transmission line in real time, and promptly discover short-circuit discharge faults, which helps to quickly take measures to repair them, reduce the impact of faults on the power system, and improve the safety and reliability of the power system.
[0019] Comprehensive detection: This method provides a new means for short-circuit discharge fault detection of transmission lines. Combined with existing current, voltage and machine vision detection methods, it can form a more comprehensive detection system and improve the detection capability of short-circuit discharge faults of transmission lines.
[0020] Furthermore, before inputting the current fused voiceprint feature into the pre-trained short-circuit recognition neural network model, the method further includes:
[0021] Collecting transmission line short-circuit discharge sound data;
[0022] Extracting Mel-spectrogram features and Mel-cepstral coefficient features of the transmission line short-circuit discharge sound data;
[0023] Constructing a fused voiceprint feature of the power transmission line short-circuit discharge sound data according to the Mel-frequency spectrum feature and the Mel-frequency cepstrum coefficient feature;
[0024] According to the fused voiceprint features of the transmission line short-circuit discharge sound data, a pre-created initial short-circuit recognition neural network model is trained to obtain a short-circuit recognition neural network model that meets the requirements.
[0025] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0026] Targeted training: By collecting the sound data of short-circuit discharge of power transmission lines, and extracting the Mel-spectrogram features and Mel-cepstral coefficient features from them to construct the fusion voiceprint features, the initial short-circuit recognition neural network model is trained. The model trained in this way can better identify the sound features of short-circuit discharge of power transmission lines, improving the targeting and accuracy of the model.
[0027] Improve recognition accuracy: Use special transmission line short-circuit discharge sound data for training, so that the short-circuit recognition neural network model can better learn the characteristic patterns of short-circuit discharge sounds, so that in practical applications it can more accurately determine whether the current fused voiceprint feature input is generated by short-circuit discharge, reducing the possibility of misjudgment.
[0028] Enhanced model adaptability: After training with a large amount of actual short-circuit discharge sound data, the model can better adapt to the short-circuit discharge sound of transmission lines under different circumstances, which improves the generalization ability and adaptability of the model, enabling it to maintain good recognition performance in the face of various complex actual scenarios.
[0029] Optimize model performance: By continuously adjusting the model parameters, the model can be better fitted to the training data, thereby improving the performance and accuracy of the model and providing more reliable support for subsequent short-circuit discharge fault detection.
[0030] Save time and cost: In actual detection, the use of a trained short-circuit identification neural network model that meets the requirements can quickly analyze and judge the collected sound information, saving the time and cost of manual analysis and improving detection efficiency.
[0031] Furthermore, the transmission line short-circuit discharge sound data includes: real transmission line short-circuit discharge sound data, simulated short-circuit discharge sound data and expanded short-circuit discharge sound data;
[0032] The real power transmission line short-circuit discharge sound data includes multiple types of real power transmission line short-circuit discharge sound data;
[0033] The simulated short-circuit discharge sound data is generated based on a mixture of different types of real power transmission line short-circuit discharge sound data;
[0034] The expanded short-circuit discharge sound data is generated based on different types of real power transmission line short-circuit discharge sound data and the simulated short-circuit discharge sound data, and is aliased with noise in a power transmission line scenario.
[0035] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0036] Data diversity: Contains real transmission line short-circuit discharge sound data, simulated short-circuit discharge sound data, and expanded short-circuit discharge sound data, making the data type more diverse. Various types of real transmission line short-circuit discharge sound data further increase the diversity of the data, which helps the model learn the short-circuit discharge sound characteristics under different circumstances and improve the generalization ability of the model.
[0037] Advantages of simulated data: The simulated short-circuit discharge sound data is generated based on a mixture of different types of real transmission line short-circuit discharge sound data, which can supplement the deficiencies that may exist in real data and increase the coverage of data. Through simulated data, the model can learn more short-circuit discharge sound patterns under different circumstances and improve its adaptability to complex situations.
[0038] Significance of extended data: Extended short-circuit discharge sound data is generated based on real data and simulated data mixed with noise in the transmission line scenario. Such extended data is closer to the actual transmission line operating environment and takes into account the impact of noise. This enables the model to better learn how to accurately identify short-circuit discharge sounds in a noisy environment during training, improving the model's anti-interference ability and accuracy in practical applications.
[0039] Improve model accuracy: Rich and diverse data types and data generation methods that are closer to actual conditions can provide the model with more comprehensive learning materials, which helps the model learn the characteristics of short-circuit discharge sounds more accurately, thereby improving the accuracy and reliability of the short-circuit identification neural network model.
[0040] Enhanced model adaptability: By using multiple types of real data, simulated data, and augmented data for training, the model can better adapt to different transmission line scenarios and noise environments, improving the adaptability and stability of the model in practical applications.
[0041] Furthermore, the extracting of Mel-spectrogram features and Mel-cepstral coefficient features of the transmission line short-circuit discharge sound data includes:
[0042] According to a preset pre-emphasis system, pre-emphasis processing is performed on the power transmission line short-circuit discharge sound data to obtain pre-emphasized power transmission line short-circuit discharge sound data;
[0043] Dividing the pre-emphasized power transmission line short-circuit discharge sound data into multiple frames of sound data according to time;
[0044] Multiply each frame of sound data by a preset window function to obtain windowed transmission line short-circuit discharge sound data;
[0045] The Mel-spectrogram features and Mel-cepstral coefficient features of the windowed transmission line short-circuit discharge sound data are extracted.
[0046] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0047] Improve signal quality: Pre-emphasize the sound data of short-circuit discharge of transmission lines through a pre-set pre-emphasis system, which can enhance the high-frequency part of the sound signal, compensate for the high-frequency attenuation of the sound during transmission, improve the quality and clarity of the signal, and facilitate subsequent feature extraction and analysis.
[0048] Advantages of frame processing: The pre-emphasized sound data is divided into multiple frames of sound data according to time, so that the signal is divided into smaller segments in time. This can better capture the time-varying characteristics of the sound signal, improve the analysis accuracy of the sound signal, and help process non-stationary signals.
[0049] Reduce spectrum leakage: Multiplying each frame of sound data by a preset window function can make the signals at both ends of the frame smoothly transition to zero, thereby reducing spectrum leakage. Spectrum leakage may lead to reduced frequency resolution and energy dispersion. Windowing can improve the accuracy of spectrum analysis.
[0050] Accurate feature extraction: After the above preprocessing, the Mel spectrum features and Mel cepstral coefficient features of the windowed transmission line short-circuit discharge sound data are extracted, which can more accurately reflect the essential characteristics of the sound signal. These features are of great significance for the recognition and classification of short-circuit discharge sounds, and help improve the accuracy and reliability of short-circuit identification.
[0051] Enhanced model performance: Accurately extracted Mel-spectrogram features and Mel-cepstral coefficient features can provide high-quality input data for the subsequent short-circuit identification neural network model, thereby improving the training effect and performance of the model, enabling it to better identify the short-circuit discharge sound of the transmission line.
[0052] Improve system adaptability: Through preprocessing and feature extraction of sound data, the system can better adapt to different transmission line short-circuit discharge sound conditions and improve the system's versatility and adaptability.
[0053] Improve fault detection capabilities: Helps to more accurately detect short-circuit discharge faults in transmission lines, identify problems in a timely manner and take appropriate measures to improve the safety and reliability of the power system.
[0054] Furthermore, the extracting of Mel-spectrogram features and Mel-cepstral coefficient features of the transmission line short-circuit discharge sound data includes:
[0055] Performing Fourier transformation on each frame of the transmission line short-circuit discharge sound data to obtain Fourier transformation data, and obtaining a power spectrum signal based on the Fourier transformation data;
[0056] Each filter in the Mel filter bank is point-multiplied with the power spectrum signal to obtain the Mel spectrum feature;
[0057] The logarithm of the power spectrum signal after the Mel filter bank is taken to obtain a logarithmic processing result, and the logarithmic processing result is subjected to discrete cosine transformation to obtain a Mel cepstral coefficient feature.
[0058] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0059] In-depth analysis of sound signals: Perform Fourier transform on each frame of sound data and obtain the power spectrum signal, which can convert the sound signal from the time domain to the frequency domain, analyze the frequency components of the sound more deeply, and provide a basis for subsequent feature extraction.
[0060] Highlight key features: The Mel-spectrogram features are obtained by performing a dot product between the Mel filter bank and the power spectrum signal. This can highlight the characteristics of the human auditory system's perception of different frequencies, making the extracted features more consistent with human auditory perception and helping to improve the accuracy of identifying short-circuit discharge sounds.
[0061] Enhance the distinguishability of features: Taking the logarithm of the power spectrum signal after the Mel filter bank can compress the dynamic range of the signal, highlight weaker signal components, and enhance the distinguishability of features.
[0062] Improve the accuracy of the model: The obtained Mel-frequency cepstral coefficient features have good noise resistance and robustness. Using them as input features can improve the accuracy and reliability of the short-circuit recognition neural network model.
[0063] Better adaptability to actual application scenarios: These feature extraction methods take into account the characteristics of human auditory perception and sound signals, so that in practical applications, they can better cope with complex environmental noise and signal changes and improve the performance of the system in actual scenarios.
[0064] Provide strong support for fault diagnosis: Accurately extracted Mel-spectrogram features and Mel-cepstral coefficient features help to more accurately identify the short-circuit discharge sound of transmission lines, provide strong support for timely detection and diagnosis of transmission line faults, and improve the safety and stability of the power system.
[0065] Furthermore, according to the Mel frequency spectrum feature and the Mel cepstral coefficient feature, a fused voiceprint feature of the transmission line short-circuit discharge sound data is constructed, including:
[0066] Selecting a real transmission line short-circuit discharge sound data from the transmission line short-circuit discharge sound data as reference sound data, and extracting a reference Mel-frequency spectrum feature and a reference Mel-frequency cepstral coefficient feature of the reference sound data;
[0067] Extracting Mel-spectrogram features and Mel-cepstral coefficient features of the same feature dimension from all short-circuit discharge sounds in the transmission line short-circuit discharge sound data;
[0068] Calculate feature similarities of the Mel-spectrogram features and Mel-cepstral coefficient features of all short-circuit discharge sounds and the benchmark Mel-spectrogram features and benchmark Mel-cepstral coefficient features of the benchmark sound data, respectively, to obtain an average feature similarity score of the Mel-spectrogram features and an average feature similarity score of the Mel-cepstral coefficient features;
[0069] Constructing a feature fusion weight corresponding to the Mel-spectrogram feature and a feature fusion weight corresponding to the Mel-cepstral coefficient feature required for fusion of voiceprint features according to the average feature similarity score of the Mel-spectrogram feature and the average feature similarity score of the Mel-cepstral coefficient feature;
[0070] A fused voiceprint feature of the power transmission line short-circuit discharge sound data is constructed based on the Mel-spectrogram feature, the feature fusion weight corresponding to the Mel-spectrogram feature, the Mel-cepstral coefficient feature, and the feature fusion weight corresponding to the Mel-cepstral coefficient feature.
[0071] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0072] Reference role of benchmark data: Select a real transmission line short-circuit discharge sound data as the benchmark sound data, and extract its features, which provides a reference standard for subsequent feature comparison and fusion, and helps to ensure the accuracy and reliability of the fused voiceprint features.
[0073] Consistency of feature dimensions: Mel-spectrogram features and Mel-cepstral coefficient features of the same feature dimensions are extracted for all short-circuit discharge sounds to ensure the consistency and comparability of the data, facilitating subsequent feature similarity calculation and fusion operations.
[0074] Feature similarity evaluation: By calculating the feature similarity of all short-circuit discharge sounds and the benchmark sound data, the average feature similarity score of the Mel-spectrogram feature and the Mel-cephalogram coefficient feature is obtained. This helps to understand the degree of similarity between different sound data and provides an important basis for constructing fused voiceprint features.
[0075] Reasonable feature fusion weight: The feature fusion weight required for fusion voiceprint features is constructed according to the average score of feature similarity, making the fusion process more scientific and reasonable. It can assign appropriate weights to Mel spectrum features and Mel cepstral coefficient features according to the importance and similarity of the features, thus improving the quality and effectiveness of fusion voiceprint features.
[0076] Constructing fused voiceprint features: Taking into account the Mel-spectrogram features, the feature fusion weights corresponding to the Mel-spectrogram features, the Mel-cepstral coefficient features, and the feature fusion weights corresponding to the Mel-cepstral coefficient features, the fused voiceprint features of the transmission line short-circuit discharge sound data are constructed. This fused voiceprint feature can make full use of the advantages of the two features, more comprehensively describe the characteristics of the short-circuit discharge sound, and improve the recognition and analysis capabilities of the short-circuit discharge sound.
[0077] Improve recognition accuracy: Compared with a single feature, the fused voiceprint feature can provide richer and more accurate information, which helps to improve the recognition accuracy of the short-circuit discharge sound of the transmission line and reduce the possibility of misjudgment and missed judgment.
[0078] Enhance system reliability: By building a fusion voiceprint feature, the reliability and stability of the short-circuit discharge sound detection system can be improved, enabling it to better adapt to different transmission line operating environments and fault conditions.
[0079] Provide stronger support for fault diagnosis: Accurate fusion of voiceprint features can provide stronger support for fault diagnosis of transmission lines, help timely discover and locate short-circuit faults, and improve the safety and operation efficiency of the power system.
[0080] Furthermore, constructing the feature fusion weights corresponding to the Mel-spectrogram features and the feature fusion weights corresponding to the Mel-cepstral coefficient features required for fusion of voiceprint features according to the feature similarity average scores of the Mel-spectrogram features and the feature similarity average scores of the Mel-cepstral coefficient features, includes:
[0081] According to the formula and Construct the feature fusion weights corresponding to the Mel-spectrogram features and the feature fusion weights corresponding to the Mel-cepstral coefficient features required for fusion voiceprint features;
[0082] Among them, α is the feature fusion weight corresponding to the Mel-spectrogram feature, β is the feature fusion weight corresponding to the Mel-cepstral coefficient feature, avg1 is the average feature similarity score of the Mel-spectrogram feature, and avg2 is the average feature similarity score of the Mel-cepstral coefficient feature.
[0083] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0084] Scientifically assign weights: By using the above formula to construct feature fusion weights based on the average feature similarity scores of the Mel-spectrogram features and the Mel-cephalogram coefficient features, the importance of the two features in the fused voiceprint features can be scientifically determined. This weight assignment method based on feature similarity makes the determination of fusion weights more objective and reasonable.
[0085] Highlighting dominant features: This method can assign greater weights to features with higher similarity based on the similarity of features, thereby highlighting features that are more advantageous in describing the sound data of power transmission line short-circuit discharge. This can improve the accuracy and effectiveness of the fused voiceprint features and better reflect the essential characteristics of the sound data.
[0086] Improve fusion effect: Reasonable feature fusion weights can better combine Mel-spectrogram features and Mel-cepstral coefficient features in the fused voiceprint features, give full play to their respective advantages, and thus improve the fusion effect. This helps to more accurately identify and analyze the short-circuit discharge sound of the transmission line and improve the accuracy of fault detection.
[0087] Enhance system adaptability: By dynamically adjusting the fusion weight according to feature similarity, the fused voiceprint feature can better adapt to different transmission line short-circuit discharge sound conditions. This enhances the adaptability and robustness of the system in the face of various practical application scenarios.
[0088] Furthermore, constructing a current fused voiceprint feature according to the current Mel-spectrogram feature and the current Mel-cepstral coefficient feature includes:
[0089] The current fused voiceprint feature is constructed according to the current Mel-spectrogram feature, the feature fusion weight corresponding to the Mel-spectrogram feature, the current Mel-cepstral coefficient feature, and the feature fusion weight corresponding to the Mel-cepstral coefficient feature.
[0090] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0091] Comprehensively characterize sound features: Combining the current Mel-spectrogram features and the current Mel-cepstral coefficient features can more comprehensively describe the characteristics of the transmission line short-circuit discharge sound. The Mel-spectrogram features reflect the frequency information of the sound, while the Mel-cepstral coefficient features contain the comprehensive information of the sound in the time domain and frequency domain. Fusion of these two features can provide a richer and more comprehensive representation of sound features.
[0092] Improve the accuracy of features: By setting corresponding feature fusion weights for the Mel spectrum features and Mel cepstral coefficient features, it is possible to reasonably weight the two features according to their importance in describing short-circuit discharge sound data. This allows the fused features to more accurately reflect the essential characteristics of the sound and reduce the errors and limitations that may be caused by a single feature.
[0093] Enhanced fault detection capability: The current fusion voiceprint feature constructed can more effectively identify the sound of short-circuit discharge in power transmission lines. This helps to improve the accuracy and reliability of fault detection, timely discover potential short-circuit faults, and reduce safety hazards in the power system.
[0094] Adapt to complex environments: Fusion of voiceprint features has stronger anti-interference ability and adaptability. In the actual transmission line operation environment, there may be various noise and interference factors. Fusion of multiple features and reasonable weight allocation can make the features more robust, better able to cope with complex environmental changes, and improve the stability of fault detection.
[0095] Furthermore, the fused voiceprint feature of the power transmission line short-circuit discharge sound data is constructed according to the Mel-spectrogram feature, the feature fusion weight corresponding to the Mel-spectrogram feature, the Mel-cepstral coefficient feature, and the feature fusion weight corresponding to the Mel-cepstral coefficient feature, including:
[0096] According to the formula fmix=α*f1+β*f2, a fusion voiceprint feature of the transmission line short-circuit discharge sound data is constructed;
[0097] Among them, fmix is the fused voiceprint feature of the transmission line short-circuit discharge sound data, α is the feature fusion weight corresponding to the Mel spectrum feature, β is the feature fusion weight corresponding to the Mel cepstral coefficient feature, f1 is the Mel spectrum feature, and f2 is the Mel cepstral coefficient feature.
[0098] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0099] Optimized feature representation: By fusing the Mel-spectrogram feature and the Mel-cepstral coefficient feature and weighting them according to their respective feature fusion weights, a more optimized fused voiceprint feature of the transmission line short-circuit discharge sound data can be obtained. This fused feature can more comprehensively and accurately reflect the characteristics of the sound, avoiding the limitations that may exist in a single feature.
[0100] Improve fault diagnosis accuracy: Fusion of voiceprint features can better capture the details and features of the sound of short-circuit discharge in power transmission lines, thereby improving the accuracy of short-circuit fault diagnosis. This helps to detect faults more promptly and accurately, reduce misjudgments and missed judgments, and improve the safety and reliability of the power system.
[0101] Enhanced robustness of features: The fused voiceprint features that take into account the feature fusion weights have stronger resistance to noise and other interference. Even in complex environments, they can maintain good performance, improving the robustness and stability of features.
[0102] One or more embodiments of this specification provide a transmission line short circuit identification device, including:
[0103] at least one processor; and,
[0104] a memory communicatively connected to the at least one processor; wherein,
[0105] The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:
[0106] Collect current sound information within the preset range of the transmission line;
[0107] Extracting current Mel-spectrogram features and current Mel-cepstral coefficient features of the current sound information;
[0108] Constructing a current fused voiceprint feature according to the current Mel-frequency spectrum feature and the current Mel-frequency cepstral coefficient feature;
[0109] The current fused voiceprint feature is input into a pre-trained short-circuit recognition neural network model, and a recognition result is output, so as to determine whether the current sound information is generated by short-circuit discharge according to the recognition result.
[0110] At least one of the above technical solutions adopted in the embodiments of this specification can achieve the following beneficial effects:
[0111] Improved accuracy: By collecting the current sound information within the preset range of the transmission line, extracting its Mel spectrum features and Mel cepstral coefficient features, constructing the current fused voiceprint features, and then inputting them into the pre-trained short-circuit recognition neural network model for recognition. This recognition method based on sound features can more accurately determine whether the sound is caused by short-circuit discharge, avoiding the noise and interference that the instantaneous value of current or the true effective value method may be affected by, as well as the high requirements for equipment accuracy of the voltage and phase angle change detection method, thereby improving the accuracy of short-circuit discharge fault detection.
[0112] Enhanced reliability: This method uses sound information to identify short circuits, which is not affected by environmental factors such as light and fog. Compared with machine vision detection methods, it has higher reliability. It can stably detect under various environmental conditions, reducing misjudgments and inaccurate detection caused by environmental factors.
[0113] Economic advantage: Compared with the voltage and phase angle change detection method, this method may not require high-precision expensive equipment, reducing the detection cost. At the same time, detection through sound information is relatively simple, which also helps to reduce maintenance and operating costs.
[0114] Good real-time performance: It can collect and analyze the sound information of the transmission line in real time, and promptly discover short-circuit discharge faults, which helps to quickly take measures to repair them, reduce the impact of faults on the power system, and improve the safety and reliability of the power system.
[0115] Comprehensive detection: This method provides a new means for short-circuit discharge fault detection of transmission lines. Combined with existing current, voltage and machine vision detection methods, it can form a more comprehensive detection system and improve the detection capability of short-circuit discharge faults of transmission lines. BRIEF DESCRIPTION OF THE DRAWINGS
[0116] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art description. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. In the drawings:
[0117] Figure 1 A schematic flow chart of a method for identifying a short circuit in a power transmission line provided for one or more embodiments of this specification;
[0118] Figure 2 A sound feature extraction flow chart provided for one or more embodiments of this specification;
[0119] Figure 3 A flow chart of a method for fusing voiceprint features provided for one or more embodiments of this specification;
[0120] Figure 4 A schematic diagram of a sub-sampling time-delay neural network structure provided for one or more embodiments of this specification;
[0121] Figure 5 A flow chart of abnormal short-circuit discharge sound detection of a power transmission line provided for one or more embodiments of this specification;
[0122] Figure 6 A schematic diagram of the structure of a power transmission line short circuit identification device provided in one or more embodiments of this specification. DETAILED DESCRIPTION
[0123] The embodiments of this specification provide a method and device for identifying a short circuit in a power transmission line.
[0124] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.
[0125] Figure 1 A schematic diagram of a process flow of a power transmission line short circuit identification method provided in one or more embodiments of this specification, the process flow can be executed by a power transmission line short circuit identification system. Certain input parameters or intermediate results in the process flow allow manual intervention and adjustment to help improve accuracy.
[0126] The method steps of the embodiment of this specification are as follows:
[0127] S102, collecting current sound information within a preset range of the power transmission line.
[0128] In the examples of this specification, the above content can be implemented by the following specific implementation scheme:
[0129] Equipment selection: You can choose a highly sensitive sound sensor to accurately capture the sound around the power transmission line. In addition, you can also consider using a sensor with waterproof, dustproof and anti-interference capabilities to adapt to complex outdoor environmental conditions.
[0130] Installation location determination: Sound sensors can be installed at appropriate locations according to the direction and preset range of the transmission line. They can be installed on transmission line towers, fixed support points near the line, or monitoring sites. During installation, it is necessary to ensure that the installation location of the sensor can cover the key areas within the preset range to improve the comprehensiveness and accuracy of sound information collection.
[0131] Data collection settings: Determine the frequency and sampling rate of sound collection. According to actual needs and sound characteristics, select the appropriate collection frequency and sampling rate to ensure that sufficient sound details can be captured. Set the data storage method. You can choose to transmit the collected sound data to the central server in real time for storage, or store it locally first and upload it to the server regularly.
[0132] S104, extracting current Mel-spectrogram features and current Mel-cepstral coefficient features of the current sound information.
[0133] In the examples of this specification, the above content can be implemented by the following specific implementation scheme:
[0134] Data preparation: Preprocess the current sound data, which may include noise removal, filtering and other operations to improve the accuracy of subsequent feature extraction.
[0135] Mel-spectrogram feature extraction: You can choose a suitable audio processing library, such as Librosa, matplotlib, etc. You can use the functions in the selected library to convert the sound signal into a time domain signal and perform frame processing. Perform a fast Fourier transform (FFT) on each frame signal to convert the time domain signal into a frequency domain signal. Then use a Mel filter bank to convert the frequency domain signal into a Mel spectrum. Finally, perform a logarithmic operation on the Mel spectrum to obtain the Mel-spectrogram feature.
[0136] Mel cepstral coefficient feature extraction: After obtaining the Mel spectrum graph features, the Mel cepstral coefficient features can be further calculated. The Mel spectrum graph can be transformed using discrete cosine transform (DCT) to obtain Mel cepstral coefficients.
[0137] S106: construct a current fused voiceprint feature according to the current Mel-spectrogram feature and the current Mel-cepstral coefficient feature.
[0138] In the examples of this specification, the above content can be implemented by the following specific implementation scheme:
[0139] Feature fusion method selection: Consider using a variety of fusion methods, such as serial fusion, weighted fusion, etc. Serial fusion is to directly connect the Mel spectrum graph feature and Mel cepstral coefficient feature into a longer feature vector; weighted fusion is to perform a weighted sum of two features according to a certain weight. You can choose the most appropriate fusion method according to the specific application scenario and needs.
[0140] Serial fusion implementation scheme: If serial fusion is selected, the current Mel-spectrogram feature and the current Mel-cepstral coefficient feature are sequentially connected to form a new feature vector. For example, if the Mel-spectrogram feature is a vector with a dimension of M and the Mel-cepstral coefficient feature is a vector with a dimension of N, then the dimension of the feature vector after serial fusion is M+N.
[0141] Weighted fusion implementation plan: For weighted fusion, first determine the weights of the Mel-spectrogram feature and the Mel-cepstral coefficient feature. The weight values can be determined by experiment or based on prior knowledge. For example, you can find the weight combination that makes the model perform best by testing on the training set. Then, multiply the current Mel-spectrogram feature by its corresponding weight, multiply the current Mel-cepstral coefficient feature by its corresponding weight, and finally add the two weighted feature vectors to obtain the fused feature vector.
[0142] S108, inputting the current fused voiceprint feature into a pre-trained short-circuit recognition neural network model, and outputting a recognition result, so as to determine whether the current sound information is generated by short-circuit discharge according to the recognition result.
[0143] In the examples of this specification, the above content can be implemented by the following specific implementation scheme:
[0144] Model preparation: Ensure that a pre-trained short circuit identification neural network model already exists and can be loaded and run in the corresponding computing environment. Check the model's parameters, architecture, and training data to ensure that it is suitable for the task of identifying short circuit discharge sounds.
[0145] Data preprocessing: Perform necessary preprocessing on the current fused voiceprint features, such as data standardization or normalization, so that the distribution of input data meets the requirements of the model.
[0146] Model loading and initialization: Use an appropriate deep learning framework (such as TDNN, TensorFlow, PyTorch, etc.) to load a pre-trained short-circuit recognition neural network model. Initialize the model's parameters and state to prepare for input data processing.
[0147] Input data: The preprocessed current fused voiceprint features are used as input data and passed to the short-circuit recognition neural network model.
[0148] Model inference: Run the short circuit identification neural network model to infer and calculate the input fusion voiceprint features. The model will analyze and judge the input data based on the patterns and features it has learned to determine whether there are characteristics of short circuit discharge.
[0149] Output recognition result: The output of the model will be a recognition result, usually expressed as a probability or classification label of short-circuit discharge or non-short-circuit discharge. The output of the model is obtained and interpreted and transformed to clearly indicate whether the current sound information is considered to be generated by short-circuit discharge.
[0150] Result judgment and application: According to the output recognition result, determine whether the current sound information is caused by short-circuit discharge. If the recognition result shows that the probability of short-circuit discharge is high or it is classified as short-circuit discharge, corresponding measures may need to be taken, such as further inspection or maintenance. If the recognition result shows that it is not short-circuit discharge, corresponding processing or monitoring can also be carried out according to the actual situation.
[0151] Furthermore, before inputting the current fused voiceprint feature into the pre-trained short-circuit identification neural network model, it is necessary to train the short-circuit identification neural network model, and the training process is as follows:
[0152] Collect transmission line short-circuit discharge sound data; extract the Mel-spectrogram features and Mel-cepstral coefficient features of the transmission line short-circuit discharge sound data; construct a fused voiceprint feature of the transmission line short-circuit discharge sound data based on the Mel-spectrogram features and the Mel-cepstral coefficient features; train a pre-created initial short-circuit recognition neural network model based on the fused voiceprint features of the transmission line short-circuit discharge sound data to obtain a short-circuit recognition neural network model that meets the requirements.
[0153] It should be noted that the above-mentioned training short circuit identification neural network model can be implemented by the following specific implementation scheme:
[0154] Collect transmission line short-circuit discharge sound data: Choose to collect sound data at the actual transmission line site or in a simulated short-circuit discharge experimental environment. The sound collection equipment can use high-quality microphones or acoustic sensors to ensure that the short-circuit discharge sound can be accurately captured. At the same time, multiple data collections can be performed to obtain sufficient diversity and representative data to cover different short-circuit situations and environmental conditions. In addition, the collected data needs to be annotated to mark whether each sound data is a short-circuit discharge sound for subsequent model training and verification.
[0155] Data preprocessing: Preprocess the collected sound data, including denoising, filtering and other operations to improve data quality.
[0156] Calculate Mel-spectrogram features: Divide the sound signal into short time frames. Perform fast Fourier transform (FFT) on each frame signal to obtain spectrum information. Use Mel filter bank to convert the spectrum into Mel spectrum. Perform logarithmic operation on the Mel spectrum to obtain Mel-spectrogram features.
[0157] Calculate the Mel-frequency cepstral coefficient features: After obtaining the Mel-frequency spectrum features, perform discrete cosine transform (DCT) on them to obtain the Mel-frequency cepstral coefficient features.
[0158] Feature fusion method selection: You can choose methods such as serial fusion or weighted fusion.
[0159] Concatenated fusion: directly concatenate the Mel-spectrogram features and the Mel-cepstral coefficient features into a longer feature vector.
[0160] Weighted fusion: Determine the weights of the Mel-spectrogram features and the Mel-cepstral coefficient features, and then perform weighted summation on them to obtain the fused features.
[0161] Model architecture design: Choose a neural network architecture suitable for the sound recognition task, such as the time-delay neural network (TDNN).
[0162] Data division: Divide the fused voiceprint feature data into training set, validation set and test set.
[0163] Training model: Use the training set data to train the initial short circuit identification neural network model. During the training process, according to the model's loss function and optimization algorithm, the model parameters are continuously adjusted to improve the model's performance.
[0164] Model evaluation and adjustment: Use the validation set data to evaluate the model during training, and adjust the model's hyperparameters such as learning rate, number of layers, number of nodes, etc. based on the evaluation results. Repeat the training and evaluation process until the model's performance on the validation set meets the expected requirements.
[0165] Model testing and verification: Use the test set data to test the final trained model and evaluate the generalization ability and accuracy of the model. If the performance of the model on the test set does not meet the requirements, further analysis of the reasons and improvements are required, which may include adjusting the data preprocessing method, increasing the amount of data, improving the model architecture, etc.
[0166] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0167] Targeted training: By collecting the sound data of short-circuit discharge of power transmission lines, and extracting the Mel-spectrogram features and Mel-cepstral coefficient features from them to construct the fusion voiceprint features, the initial short-circuit recognition neural network model is trained. The model trained in this way can better identify the sound features of short-circuit discharge of power transmission lines, improving the targeting and accuracy of the model.
[0168] Improve recognition accuracy: Use special transmission line short-circuit discharge sound data for training, so that the short-circuit recognition neural network model can better learn the characteristic patterns of short-circuit discharge sounds, so that in practical applications it can more accurately determine whether the current fused voiceprint feature input is generated by short-circuit discharge, reducing the possibility of misjudgment.
[0169] Enhanced model adaptability: After training with a large amount of actual short-circuit discharge sound data, the model can better adapt to the short-circuit discharge sound of transmission lines under different circumstances, which improves the generalization ability and adaptability of the model, enabling it to maintain good recognition performance in the face of various complex actual scenarios.
[0170] Optimize model performance: By continuously adjusting the model parameters, the model can be better fitted to the training data, thereby improving the performance and accuracy of the model and providing more reliable support for subsequent short-circuit discharge fault detection.
[0171] Save time and cost: In actual detection, the use of a trained short-circuit identification neural network model that meets the requirements can quickly analyze and judge the collected sound information, saving the time and cost of manual analysis and improving detection efficiency.
[0172] Furthermore, the transmission line short-circuit discharge sound data includes: real transmission line short-circuit discharge sound data, simulated short-circuit discharge sound data and extended short-circuit discharge sound data; the real transmission line short-circuit discharge sound data includes multiple types of real transmission line short-circuit discharge sound data; the simulated short-circuit discharge sound data is generated based on a mixture of different types of real transmission line short-circuit discharge sound data; the extended short-circuit discharge sound data is generated based on different types of real transmission line short-circuit discharge sound data and the simulated short-circuit discharge sound data, mixed with the noise in the transmission line scene.
[0173] It should be noted that the above-mentioned transmission line short-circuit discharge sound data can be implemented through the following specific implementation scheme:
[0174] The sound data of real transmission line short-circuit discharge are as follows:
[0175] Determine the scope of data collection: Cover various types of real transmission line short-circuit discharge conditions, such as different voltage levels, different line materials, different fault types, etc.
[0176] Choose a suitable collection location: Collect sound data at the actual transmission line site to ensure that the real short-circuit discharge sound can be obtained.
[0177] Use professional sound collection equipment: such as highly sensitive microphones and audio recording equipment to ensure the quality and accuracy of the data.
[0178] Classify and annotate the collected data: classify according to the type of short-circuit discharge and annotate the relevant information in detail for subsequent analysis and use.
[0179] The data for generating simulated short-circuit discharge sound are as follows:
[0180] Analyze the characteristics of different types of real transmission line short-circuit discharge sound data: including frequency, amplitude, duration and other aspects.
[0181] Use sound synthesis technology or related software: according to the characteristics of real data, different types of real transmission line short-circuit discharge sound data are mixed to generate simulated short-circuit discharge sound data.
[0182] Verify and adjust the generated simulation data: Ensure the rationality and accuracy of the simulation data by comparing it with the real data.
[0183] The data for generating the extended short-circuit discharge sound are as follows:
[0184] Different types of real transmission line short-circuit discharge sound data and generated simulated short-circuit discharge sound data are selected.
[0185] Collect noise data in transmission line scenarios: These noises can include natural environmental noise (such as wind and rain), equipment operation noise, etc.
[0186] Using sound processing technology: the selected real and simulated short-circuit discharge sound data are mixed with the noise data to generate expanded short-circuit discharge sound data.
[0187] Perform quality checks on augmented data: Ensure the availability and authenticity of the data to avoid excessive distortion or inconsistency with actual conditions.
[0188] Regarding data management and storage:
[0189] Establish a database: uniformly manage and store the collected real transmission line short-circuit discharge sound data, generated simulated short-circuit discharge sound data and expanded short-circuit discharge sound data.
[0190] Data labeling and classification: Detailed labeling and classification of data in the database to facilitate subsequent query and use.
[0191] Back up data regularly: to prevent data loss or corruption.
[0192] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0193] Data diversity: Contains real transmission line short-circuit discharge sound data, simulated short-circuit discharge sound data, and expanded short-circuit discharge sound data, making the data type more diverse. Various types of real transmission line short-circuit discharge sound data further increase the diversity of the data, which helps the model learn the short-circuit discharge sound characteristics under different circumstances and improve the generalization ability of the model.
[0194] Advantages of simulated data: The simulated short-circuit discharge sound data is generated based on a mixture of different types of real transmission line short-circuit discharge sound data, which can supplement the deficiencies that may exist in real data and increase the coverage of data. Through simulated data, the model can learn more short-circuit discharge sound patterns under different circumstances and improve its adaptability to complex situations.
[0195] Significance of extended data: Extended short-circuit discharge sound data is generated based on real data and simulated data mixed with noise in the transmission line scenario. Such extended data is closer to the actual transmission line operating environment and takes into account the impact of noise. This enables the model to better learn how to accurately identify short-circuit discharge sounds in a noisy environment during training, improving the model's anti-interference ability and accuracy in practical applications.
[0196] Improve model accuracy: Rich and diverse data types and data generation methods that are closer to actual conditions can provide the model with more comprehensive learning materials, which helps the model learn the characteristics of short-circuit discharge sounds more accurately, thereby improving the accuracy and reliability of the short-circuit identification neural network model.
[0197] Enhanced model adaptability: By using multiple types of real data, simulated data, and augmented data for training, the model can better adapt to different transmission line scenarios and noise environments, improving the adaptability and stability of the model in practical applications.
[0198] Furthermore, when extracting the Mel-spectrogram features and Mel-cepstral coefficient features of the transmission line short-circuit discharge sound data, the transmission line short-circuit discharge sound data can be pre-emphasized according to a preset pre-emphasis system to obtain pre-emphasized transmission line short-circuit discharge sound data; the pre-emphasized transmission line short-circuit discharge sound data can be divided into multiple frames of sound data according to time; each frame of sound data is multiplied by a preset window function to obtain windowed transmission line short-circuit discharge sound data; and the Mel-spectrogram features and Mel-cepstral coefficient features of the windowed transmission line short-circuit discharge sound data can be extracted.
[0199] It should be noted that the Mel-spectrogram features and Mel-cepstral coefficient features of the above-mentioned extraction of transmission line short-circuit discharge sound data can be implemented through the following specific implementation scheme:
[0200] The pre-emphasis process is as follows:
[0201] Determine the parameters of the pre-emphasis system: According to the pre-settings, the value of the pre-emphasis coefficient is clear.
[0202] Processing of transmission line short-circuit discharge sound data: For each sound data sample point, pre-emphasis processing is performed using the following formula: y[n] = x[n]-α*x[n-1], where x[n] is the original sound data, x[n-1] is the previous sample point of the current sample point x[n], y[n] is the pre-emphasized sound data, α is the pre-emphasis coefficient, and α can be 0.97. n represents the index of the sample point.
[0203] The sound data is framed as follows:
[0204] Determine the frame length and frame shift: Select an appropriate frame length (e.g., 20-40 milliseconds) and frame shift (e.g., half the frame length) based on actual needs and the characteristics of the sound signal.
[0205] The pre-emphasized transmission line short-circuit discharge sound data is divided into multiple frames according to time: the sound data is segmented into frames by taking the determined frame length and frame shift as parameters.
[0206] The windowing process is as follows:
[0207] Select window function: Common window functions include Hamming window, Hanning window, etc. Select the appropriate window function according to the actual situation.
[0208] Windowing is performed on each frame of sound data: the selected window function is applied to each frame of sound data, that is, the frame data is multiplied by the window function value point by point to obtain the windowed transmission line short-circuit discharge sound data.
[0209] The extracted features are as follows:
[0210] Calculate Mel-spectrogram features: Perform fast Fourier transform (FFT) on the windowed sound data to obtain spectrum information. Use a Mel filter bank to convert the spectrum into a Mel-spectrogram. Perform logarithmic operation on the Mel-spectrogram to obtain Mel-spectrogram features.
[0211] Calculate the Mel-frequency cepstral coefficient features: After obtaining the Mel-frequency spectrum features, perform discrete cosine transform (DCT) on them to obtain the Mel-frequency cepstral coefficient features.
[0212] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0213] Improve signal quality: Pre-emphasize the sound data of short-circuit discharge of transmission lines through a pre-set pre-emphasis system, which can enhance the high-frequency part of the sound signal, compensate for the high-frequency attenuation of the sound during transmission, improve the quality and clarity of the signal, and facilitate subsequent feature extraction and analysis.
[0214] Advantages of frame processing: The pre-emphasized sound data is divided into multiple frames of sound data according to time, so that the signal is divided into smaller segments in time. This can better capture the time-varying characteristics of the sound signal, improve the analysis accuracy of the sound signal, and help process non-stationary signals.
[0215] Reduce spectrum leakage: Multiplying each frame of sound data by a preset window function can make the signals at both ends of the frame smoothly transition to zero, thereby reducing spectrum leakage. Spectrum leakage may lead to reduced frequency resolution and energy dispersion. Windowing can improve the accuracy of spectrum analysis.
[0216] Accurate feature extraction: After the above preprocessing, the Mel spectrum features and Mel cepstral coefficient features of the windowed transmission line short-circuit discharge sound data are extracted, which can more accurately reflect the essential characteristics of the sound signal. These features are of great significance for the recognition and classification of short-circuit discharge sounds, and help improve the accuracy and reliability of short-circuit identification.
[0217] Enhanced model performance: Accurately extracted Mel-spectrogram features and Mel-cepstral coefficient features can provide high-quality input data for the subsequent short-circuit identification neural network model, thereby improving the training effect and performance of the model, enabling it to better identify the short-circuit discharge sound of the transmission line.
[0218] Improve system adaptability: Through preprocessing and feature extraction of sound data, the system can better adapt to different transmission line short-circuit discharge sound conditions and improve the system's versatility and adaptability.
[0219] Improve fault detection capabilities: Helps to more accurately detect short-circuit discharge faults in transmission lines, identify problems in a timely manner and take appropriate measures to improve the safety and reliability of the power system.
[0220] Furthermore, with regard to extracting the Mel-spectrogram features and Mel-cepstral coefficient features of the transmission line short-circuit discharge sound data, specifically: performing Fourier transform on each frame of sound data in the transmission line short-circuit discharge sound data to obtain Fourier transform data, and obtaining a power spectrum signal based on the Fourier transform data; performing dot multiplication on each filter in the Mel filter group and the power spectrum signal to obtain Mel-spectrogram features; taking the logarithm of the power spectrum signal passing through the Mel filter group to obtain a logarithmic processing result, and performing discrete cosine transform on the logarithmic processing result to obtain Mel-cepstral coefficient features.
[0221] It should be noted that the following specific implementation scheme can be used to extract the Mel-spectrogram features and Mel-cepstral coefficient features of the transmission line short-circuit discharge sound data:
[0222] Sound data preprocessing: Ensure that the quality and format of transmission line short-circuit discharge sound data meet the requirements of subsequent processing.
[0223] Framing: Determine the appropriate frame length and frame shift. Divide the sound data into multiple frames of sound data.
[0224] Fourier transform: Perform fast Fourier transform (FFT) on each frame of sound data to obtain Fourier transform data.
[0225] Calculate the power spectrum signal: Take the square of the modulus of the Fourier transform data to obtain the power spectrum signal.
[0226] Calculate Mel-spectrogram features: Design a Mel filter bank and determine the number of filters, center frequency, bandwidth and other parameters. Perform a dot multiplication of each filter in the Mel filter bank with the power spectrum signal. Sum the dot multiplication results to obtain the Mel-spectrogram features.
[0227] Calculate the Mel cepstral coefficient feature: Take the logarithm of the power spectrum signal after the Mel filter bank to obtain the logarithmic processing result. Perform discrete cosine transform (DCT) on the logarithmic processing result.
[0228] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0229] In-depth analysis of sound signals: Perform Fourier transform on each frame of sound data and obtain the power spectrum signal, which can convert the sound signal from the time domain to the frequency domain, analyze the frequency components of the sound more deeply, and provide a basis for subsequent feature extraction.
[0230] Highlight key features: The Mel-spectrogram features are obtained by performing a dot product between the Mel filter bank and the power spectrum signal. This can highlight the characteristics of the human auditory system's perception of different frequencies, making the extracted features more consistent with human auditory perception and helping to improve the accuracy of identifying short-circuit discharge sounds.
[0231] Enhance the distinguishability of features: Taking the logarithm of the power spectrum signal after the Mel filter bank can compress the dynamic range of the signal, highlight weaker signal components, and enhance the distinguishability of features.
[0232] Improve the accuracy of the model: The obtained Mel-frequency cepstral coefficient features have good noise resistance and robustness. Using them as input features can improve the accuracy and reliability of the short-circuit recognition neural network model.
[0233] Better adaptability to actual application scenarios: These feature extraction methods take into account the characteristics of human auditory perception and sound signals, so that in practical applications, they can better cope with complex environmental noise and signal changes and improve the performance of the system in actual scenarios.
[0234] Provide strong support for fault diagnosis: Accurately extracted Mel-spectrogram features and Mel-cepstral coefficient features help to more accurately identify the short-circuit discharge sound of transmission lines, provide strong support for timely detection and diagnosis of transmission line faults, and improve the safety and stability of the power system.
[0235] Furthermore, when constructing the fused voiceprint features of the power transmission line short-circuit discharge sound data according to the Mel-spectrogram features and the Mel-cepstral coefficient features, a real power transmission line short-circuit discharge sound data can be selected from the power transmission line short-circuit discharge sound data as the benchmark sound data, and the benchmark Mel-spectrogram features and the benchmark Mel-cepstral coefficient features of the benchmark sound data can be extracted; the Mel-spectrogram features and the Mel-cepstral coefficient features of the same feature dimensions are extracted for all short-circuit discharge sounds in the power transmission line short-circuit discharge sound data; the Mel-spectrogram features and the Mel-cepstral coefficient features of all short-circuit discharge sounds are compared with the benchmark Mel-spectrogram features and the benchmark Mel-cepstral coefficient features of the benchmark sound data. The cepstral coefficient features are respectively calculated with feature similarity to obtain the average feature similarity score of the mel-spectrogram features and the average feature similarity score of the mel-cepstral coefficient features; according to the average feature similarity score of the mel-spectrogram features and the average feature similarity score of the mel-cepstral coefficient features, the feature fusion weights corresponding to the mel-spectrogram features and the feature fusion weights corresponding to the mel-cepstral coefficient features required for the fused voiceprint features are constructed; according to the mel-spectrogram features, the feature fusion weights corresponding to the mel-spectrogram features, the mel-cepstral coefficient features, and the feature fusion weights corresponding to the mel-cepstral coefficient features, the fused voiceprint features of the transmission line short-circuit discharge sound data are constructed.
[0236] It should be noted that the following specific implementation scheme can be used to construct the fused voiceprint features of the transmission line short-circuit discharge sound data:
[0237] Select benchmark sound data and extract features: From the transmission line short-circuit discharge sound data, select a representative real transmission line short-circuit discharge sound data as the benchmark sound data. According to the previously mentioned method, the benchmark sound data is processed, including pre-emphasis, framing, windowing, and then Fourier transform, power spectrum signal calculation, and then the benchmark Mel spectrum graph feature is obtained through the Mel filter group. The power spectrum signal after the Mel filter group is logarithmic and discrete cosine transform is performed to obtain the benchmark Mel cepstrum coefficient feature.
[0238] Extracting features of all short-circuit discharge sounds: For all short-circuit discharge sounds in the transmission line short-circuit discharge sound data, the same processing flow as the benchmark sound data is adopted to extract Mel-spectrogram features and Mel-cepstral coefficient features of the same feature dimension.
[0239] Calculate feature similarity: Select an appropriate feature similarity calculation method, such as cosine similarity, Euclidean distance, etc. Calculate the similarity of the Mel-spectrogram features of all short-circuit discharge sounds with the benchmark Mel-spectrogram features of the benchmark sound data, and average all similarity values to obtain the average feature similarity score of the Mel-spectrogram features. Similarly, calculate the similarity of the Mel-cepstral coefficient features of all short-circuit discharge sounds with the benchmark Mel-cepstral coefficient features of the benchmark sound data, and average them to obtain the average feature similarity score of the Mel-cepstral coefficient features.
[0240] Constructing feature fusion weights: Based on the average feature similarity scores of the Mel-spectrogram features, the feature fusion weights corresponding to the Mel-spectrogram features are constructed through certain rules or algorithms. For example, the average similarity scores can be normalized and used as feature fusion weights. A similar method is used to construct feature fusion weights corresponding to the Mel-cepstral coefficient features based on the average feature similarity scores of the Mel-cepstral coefficient features.
[0241] Construct fused voiceprint features: Multiply the extracted Mel-spectrogram features of all short-circuit discharge sounds by their corresponding feature fusion weights to obtain weighted Mel-spectrogram features. Perform the same operation on the Mel-cepstral coefficient features of all short-circuit discharge sounds to obtain weighted Mel-cepstral coefficient features. Fuse the weighted Mel-spectrogram features and the weighted Mel-cepstral coefficient features to construct the fused voiceprint features of the transmission line short-circuit discharge sound data. The fusion method can be selected according to actual needs, such as simple splicing or more complex fusion operations.
[0242] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0243] Reference role of benchmark data: Select a real transmission line short-circuit discharge sound data as the benchmark sound data, and extract its features, which provides a reference standard for subsequent feature comparison and fusion, and helps to ensure the accuracy and reliability of the fused voiceprint features.
[0244] Consistency of feature dimensions: Mel-spectrogram features and Mel-cepstral coefficient features of the same feature dimensions are extracted for all short-circuit discharge sounds to ensure the consistency and comparability of the data, facilitating subsequent feature similarity calculation and fusion operations.
[0245] Feature similarity evaluation: By calculating the feature similarity of all short-circuit discharge sounds and the benchmark sound data, the average feature similarity score of the Mel-spectrogram feature and the Mel-cephalogram coefficient feature is obtained. This helps to understand the degree of similarity between different sound data and provides an important basis for constructing fused voiceprint features.
[0246] Reasonable feature fusion weight: The feature fusion weight required for fusion voiceprint features is constructed according to the average score of feature similarity, making the fusion process more scientific and reasonable. It can assign appropriate weights to Mel spectrum features and Mel cepstral coefficient features according to the importance and similarity of the features, thus improving the quality and effectiveness of fusion voiceprint features.
[0247] Constructing fused voiceprint features: Taking into account the Mel-spectrogram features, the feature fusion weights corresponding to the Mel-spectrogram features, the Mel-cepstral coefficient features, and the feature fusion weights corresponding to the Mel-cepstral coefficient features, the fused voiceprint features of the transmission line short-circuit discharge sound data are constructed. This fused voiceprint feature can make full use of the advantages of the two features, more comprehensively describe the characteristics of the short-circuit discharge sound, and improve the recognition and analysis capabilities of the short-circuit discharge sound.
[0248] Improve recognition accuracy: Compared with a single feature, the fused voiceprint feature can provide richer and more accurate information, which helps to improve the recognition accuracy of the short-circuit discharge sound of the transmission line and reduce the possibility of misjudgment and missed judgment.
[0249] Enhance system reliability: By building a fusion voiceprint feature, the reliability and stability of the short-circuit discharge sound detection system can be improved, enabling it to better adapt to different transmission line operating environments and fault conditions.
[0250] Provide stronger support for fault diagnosis: Accurate fusion of voiceprint features can provide stronger support for fault diagnosis of transmission lines, help timely discover and locate short-circuit faults, and improve the safety and operation efficiency of the power system.
[0251] Furthermore, when constructing the feature fusion weights corresponding to the Mel-spectrogram features and the Mel-cepstral coefficient features required for fusion voiceprint features based on the feature similarity average scores of the Mel-spectrogram features and the feature similarity average scores of the Mel-cepstral coefficient features, the formula and Construct the feature fusion weights corresponding to the Mel-spectrogram features and the feature fusion weights corresponding to the Mel-cepstral coefficient features required for fusion voiceprint features, where α is the feature fusion weight corresponding to the Mel-spectrogram features, β is the feature fusion weight corresponding to the Mel-cepstral coefficient features, avg1 is the average feature similarity score of the Mel-spectrogram features, and avg2 is the average feature similarity score of the Mel-cepstral coefficient features.
[0252] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0253] Scientifically assign weights: By using the above formula to construct feature fusion weights based on the average feature similarity scores of the Mel-spectrogram features and the Mel-cephalogram coefficient features, the importance of the two features in the fused voiceprint features can be scientifically determined. This weight assignment method based on feature similarity makes the determination of fusion weights more objective and reasonable.
[0254] Highlighting dominant features: This method can assign greater weights to features with higher similarity based on the similarity of features, thereby highlighting features that are more advantageous in describing the sound data of power transmission line short-circuit discharge. This can improve the accuracy and effectiveness of the fused voiceprint features and better reflect the essential characteristics of the sound data.
[0255] Improve fusion effect: Reasonable feature fusion weights can better combine Mel-spectrogram features and Mel-cepstral coefficient features in the fused voiceprint features, give full play to their respective advantages, and thus improve the fusion effect. This helps to more accurately identify and analyze the short-circuit discharge sound of the transmission line and improve the accuracy of fault detection.
[0256] Enhance system adaptability: By dynamically adjusting the fusion weight according to feature similarity, the fused voiceprint feature can better adapt to different transmission line short-circuit discharge sound conditions. This enhances the adaptability and robustness of the system in the face of various practical application scenarios.
[0257] Furthermore, when constructing the fused voiceprint feature of the transmission line short-circuit discharge sound data according to the Mel-spectrogram feature, the feature fusion weight corresponding to the Mel-spectrogram feature, the Mel-cepstral coefficient feature, and the feature fusion weight corresponding to the Mel-cepstral coefficient feature, the fused voiceprint feature of the transmission line short-circuit discharge sound data can be constructed according to the formula fmix=α*f1+β*f2, wherein fmix is the fused voiceprint feature of the transmission line short-circuit discharge sound data, α is the feature fusion weight corresponding to the Mel-spectrogram feature, β is the feature fusion weight corresponding to the Mel-cepstral coefficient feature, f1 is the Mel-spectrogram feature, and f2 is the Mel-cepstral coefficient feature.
[0258] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0259] Optimized feature representation: By fusing the Mel-spectrogram feature and the Mel-cepstral coefficient feature and weighting them according to their respective feature fusion weights, a more optimized fused voiceprint feature of the transmission line short-circuit discharge sound data can be obtained. This fused feature can more comprehensively and accurately reflect the characteristics of the sound, avoiding the limitations that may exist in a single feature.
[0260] Improve fault diagnosis accuracy: Fusion of voiceprint features can better capture the details and features of the sound of short-circuit discharge in power transmission lines, thereby improving the accuracy of short-circuit fault diagnosis. This helps to detect faults more promptly and accurately, reduce misjudgments and missed judgments, and improve the safety and reliability of the power system.
[0261] Enhanced robustness of features: The fused voiceprint features that take into account the feature fusion weights have stronger resistance to noise and other interference. Even in complex environments, they can maintain good performance, improving the robustness and stability of features.
[0262] Further, in S106, when constructing the current fused voiceprint feature according to the current mel-spectrogram feature and the current mel-cepstral coefficient feature, the current fused voiceprint feature can be constructed according to the current mel-spectrogram feature, the feature fusion weight corresponding to the mel-spectrogram feature, the current mel-cepstral coefficient feature, and the feature fusion weight corresponding to the mel-cepstral coefficient feature. That is, the current fused voiceprint feature can be constructed according to the above formula fmix=α*f1+β*f2, where f1 is the current mel-spectrogram feature and f2 is the current mel-cepstral coefficient feature.
[0263] It should be noted that the embodiments of this specification have the following beneficial effects regarding the above content:
[0264] Comprehensively characterize sound features: Combining the current Mel-spectrogram features and the current Mel-cepstral coefficient features can more comprehensively describe the characteristics of the transmission line short-circuit discharge sound. The Mel-spectrogram features reflect the frequency information of the sound, while the Mel-cepstral coefficient features contain the comprehensive information of the sound in the time domain and frequency domain. Fusion of these two features can provide a richer and more comprehensive representation of sound features.
[0265] Improve the accuracy of features: By setting corresponding feature fusion weights for the Mel spectrum features and Mel cepstral coefficient features, it is possible to reasonably weight the two features according to their importance in describing short-circuit discharge sound data. This allows the fused features to more accurately reflect the essential characteristics of the sound and reduce the errors and limitations that may be caused by a single feature.
[0266] Enhanced fault detection capability: The current fusion voiceprint feature constructed can more effectively identify the sound of short-circuit discharge in power transmission lines. This helps to improve the accuracy and reliability of fault detection, timely discover potential short-circuit faults, and reduce safety hazards in the power system.
[0267] Adapt to complex environments: Fusion of voiceprint features has stronger anti-interference ability and adaptability. In the actual transmission line operation environment, there may be various noise and interference factors. Fusion of multiple features and reasonable weight allocation can make the features more robust, better able to cope with complex environmental changes, and improve the stability of fault detection.
[0268] It should be noted that due to the dispersion and uncertainty of abnormal short-circuit discharge, many detection methods have limitations and cannot quickly and timely capture the occurrence of such hidden dangers. Most of the existing technologies for abnormal short-circuit fault detection use the characteristics of abnormal short-circuit discharge, which is usually accompanied by sound, light, heat, etc. There have been some studies, such as using the instantaneous value of current or the true effective value method as the basis for short-circuit judgment, detecting the voltage and phase angle changes during short-circuit, and based on machine vision detection methods.
[0269] This application proposes a method for detecting abnormal short-circuit discharge sound of power transmission lines based on fused voiceprint features. First, a database of abnormal short-circuit discharge sound of power transmission lines is constructed based on the characteristics of collected short-circuit discharge sound and data enhancement and expansion methods; then, a short-circuit discharge reference audio is selected based on this database, and Mel-spectrogram features and Mel-cepstral coefficient features are extracted for the reference audio and other short-circuit discharge audio respectively, and the similarity score mean of the corresponding features is obtained, and the fusion weight parameter is calculated based on this feature mean for fusion voiceprint feature extraction; finally, a time-delay neural network short-circuit discharge detection model iteratively trained with fused voiceprint features is used to achieve continuous and effective detection of abnormal short-circuit discharge of power transmission lines through voiceprint features of sound, thereby improving the safety level of power grid equipment operation.
[0270] The following is a specific implementation plan:
[0271] 1. Construction of transmission line short-circuit discharge sound database
[0272] Although short-circuit discharges in transmission lines do occur, they are not very frequent, so it is difficult to collect data from transmission towers. Therefore, the construction of the database is planned to start from the following three aspects.
[0273] (1) Collect the real short-circuit discharge sound of the transmission line. This part is based on the tower-mounted equipment to continuously collect the ambient sound of the transmission line and analyze it. The sound is uploaded to the storage server in a file every 10 seconds. Then, the collected audio is annotated by multiple people. A data screening mechanism is established based on the annotated file to ensure data quality. The specific method is as follows: Based on multiple annotated files, obtain whether the transmission line hidden danger sound category is marked as 1. Marking 1 means that the current audio does contain clear hidden danger sounds of this type. Then, a pre-set score is given to the audio file. All annotated files are traversed in turn to obtain the total score of the audio. When the score exceeds the screening threshold, it is considered that the audio does belong to the current hidden danger category.
[0274] (2) Analyze the characteristics of short-circuit discharge sound and simulate the short-circuit discharge sound. Generally, there are two situations of short-circuit discharge sound. One is that the wire short-circuit suddenly produces a popping sound; the other is that a low-frequency sound such as corona discharge appears first, and then a popping sound is generated as the discharge continues. The number of popping sounds is random and uncertain. Therefore, based on the characteristics of short-circuit discharge sound, low-frequency sounds such as corona discharge are intercepted from the collected real short-circuit discharge sounds, or corona discharge sounds are collected, and popping sounds are collected at the same time, and short-circuit discharge simulation sound data is generated by random combination; in addition, similar sound sources are placed in real application scenarios, and then the sounds emitted are collected using equipment. The collected sounds are further adjusted according to the average decibel value of the short-circuit discharge sound collected in the real scenario to simulate the target sound.
[0275] (3) A data expansion method suitable for short-circuit discharge sound is proposed, which specifically includes random speed perturbation, with the speed perturbation value ranging from 0.9 to 1.1. The audio is randomly perturbed with a perturbation value within the range, and the audio speed will become faster or slower according to the adjustment; random displacement perturbation, randomly moving the audio forward or backward by a few milliseconds along the time axis, and setting the movement value range between -5ms and 5ms; background noise aliasing, aliasing the sound data collected and simulated with the noise that frequently appears in the transmission line scene, increasing the difficulty of the data and improving the generalization of the model. The specific approach is as follows: first, analyze the sound type and sound decibel value around the transmission line to select the background noise, and then randomly select the target audio to be enhanced and the number of background noise types and the corresponding background noise audio. Next, randomly select the corresponding time point as the starting position of the target audio to be enhanced, and perform audio aliasing operations in sequence. Finally, the aliased audio is regularized to 10s in length to ensure data format consistency.
[0276] In addition, we focus on collecting and marking other types of sounds around transmission lines, which are mainly used as negative samples for model learning to further improve the generalization of the model. All audio in the database is in single-channel format, with a duration of 10 seconds and a sampling rate of 44100 Hz.
[0277] 2. Sound data processing and feature extraction
[0278] Figure 2 The sound feature extraction flow chart provided in the embodiment of this specification, where the data processing does not include noise reduction, in order to be better applicable to real application scenarios, the noise part in the original sound data is retained, which can effectively improve the robustness of the model and its applicability in real complex scenarios; the main pre-emphasis operation is to increase the high-frequency resolution of the speech. Assuming that the nth sampling point of the input signal is x[n], the pre-emphasis formula is as follows:
[0279] y[n]=x[n]-αx[n-1],α=0.97(α∈(0.9,1));
[0280] Where α is the pre-emphasis coefficient. This pre-processing compensates for the loss of the high-frequency part of the sound signal during the propagation process to a certain extent and protects the information of the sound channel. The sound signal has a short-term stable characteristic. In order to facilitate the subsequent Fourier transform, it is necessary to perform frame windowing. The specific method is to divide the sound signal into several segments according to time, and each segment becomes a frame. After comprehensive analysis, the frame length is set to 25ms, the frame shift is set to 15ms, and there is an overlap of 10ms between the previous and next frames to ensure a smooth transition between the two frames to prevent information leakage; to ensure that the cut signal has global continuity and facilitate subsequent frequency domain analysis, each frame signal is multiplied by a specific window function. This step is called windowing.
[0281] Mel-spectrogram feature extraction:
[0282] After the above-mentioned pre-emphasis, frame division and windowing processing steps, Fourier transform is performed on each frame respectively, and the amplitude of the data after Fourier transform is taken, squared and then divided by the set Fourier points to obtain the power spectrum.
[0283] Finally, the above data is passed through the Mel filter bank. The starting point of each filter in the Mel filter bank is at the midpoint of the previous filter. The number of filters is set to 64 here. In actual operation, each filter is point-multiplied with the power spectrum, and the result is the energy on the frequency band, which is the Mel spectrum feature (MelSpectrogram) described in this article.
[0284] MFCC feature extraction:
[0285] The first half of the MFCC extraction process is the same as the Mel-spectrogram feature extraction process. The second half is to first take the logarithm of the signal passing through the Mel filter bank, then perform discrete cosine transform, and finally calculate the dynamic features to obtain the Mel-spectrogram coefficient feature (MFCC).
[0286] 3. Constructing fusion voiceprint features
[0287] A method for extracting voiceprint features is proposed. Figure 3 This is a flow chart of the method for fusion voiceprint features of the embodiment of this specification. The fusion voiceprint model uses two underlying basic voiceprint features, one is the Mel-spectrogram feature (Mel-spectrogram), and the other is the Mel-cepstrum coefficient feature (MFCC). The two features are combined into a new fusion voiceprint feature through a specific method. Mel-spectrogram is a time-frequency representation of a sound signal, which describes the distribution of the signal's energy in time and Mel-scale frequency, where the time dimension corresponds to the frame index and the frequency dimension represents the Mel frequency band. It can capture the frequency domain content and time dynamics of the signal, and convert the representation of the sound signal into a more perceptible domain, thereby improving the performance of the algorithm. MFCC is a cepstrum coefficient calculated on the Mel scale, which effectively represents the short-time energy and frequency distribution of the sound signal. At the same time, MFCC has good robustness to noise and deformation, and can maintain good recognition performance in noisy environments. The fusion voiceprint feature combines the respective advantages of the two sound features to further improve the performance of the short-circuit discharge sound detection algorithm.
[0288] Specific steps to construct fusion voiceprint features:
[0289] Firstly, the constructed short-circuit discharge sound database is divided into training set and test set according to a specific ratio. The Mel-spectrogram features and Mel-cepstral coefficient features are extracted according to the above steps. During the feature extraction process, the parameters in the extraction process are controlled to ensure that the final feature dimensions of the two different features are consistent.
[0290] Then, a high-quality short-circuit discharge sound without background noise is selected to calculate the Mel-spectrogram features and MFCC features. Taking this as a benchmark, the Mel-spectrogram features and MFCC features of the same feature dimensions are extracted for all short-circuit discharge sounds in the database with the same parameter settings. Finally, the similarity of all features with the benchmark features is calculated and the average is taken to obtain the average similarity score of different types of acoustic features.
[0291] Next, the feature fusion weights required for fusion voiceprint features are constructed based on the calculated average similarity scores of Mel-spectrogram features and MFCC features. Assume that the average similarity score between all short-circuit discharge sound Mel-spectrogram features in the database and the benchmark Mel-spectrogram features is avg1, and the average similarity score of MFCC features is avg2. The fusion weights corresponding to the two features are α and β respectively. The specific formula is as follows:
[0292]
[0293] It should be noted that the fusion weight parameters involved in the fusion feature extraction method will change with the accumulation of data, so when the data accumulates to a certain extent, the fusion weight parameters need to be updated. After obtaining the two feature weight parameters, the fusion voiceprint feature can be constructed. At this time, f1 represents the Mel-spectrogram feature extracted from the audio data, and f2 represents the MFCC feature extracted from the audio data. Then the fusion voiceprint feature of the audio can be expressed as:
[0294] fmix=α*f1+β*f2;
[0295] Where fmix represents the fused voiceprint feature. The fused voiceprint feature extraction method is used to extract the fused voiceprint feature from all data, and then the TDNN short-circuit discharge sound detection model is trained. This model is used as the final model for short-circuit discharge sound detection. The fusion weight parameters involved in this fusion feature extraction method will change with the accumulation of data, so when the data accumulates to a certain extent, the fusion weight parameters need to be updated.
[0296] 4. Time Delay Neural Network (TDNN) Construction
[0297] Time-delay neural network (TDNN) is a feedforward neural network with the characteristics of time delay and can process time series data. TDNN is composed of convolutional layer, batch normalization layer, and fully connected layer. Relu activation function is used to add nonlinear characteristics to the model to ensure that the model has sufficient feature expression ability. The difference between TDNN and the model structure that only uses one frame of features is that it can contain the time series information of multiple frames of data. In this way, the time-delay neural network is equivalent to delaying the weight, and the overall weight is greatly reduced and easy to train. For example, to extract the feature expression with 5 frames of context, a total of 11 frames, generally speaking, the 11 frames of features will be directly spliced together to form a 11*D (D is the feature dimension of a frame) feature, and then the model will learn the 11*D feature map. TDNN processing is different. TDNN will process the time series context narrower than 11 frames in the initial layer, and then extract the 11-frame feature context feature by sending it to a deeper network.
[0298] from Figure 4 It can be seen from the schematic diagram of the sub-sampling time delay neural network structure that the time resolution of the bottom layer is 5, while the time resolution of the top layer is 23. The TDNN input in this method is the sound feature of Tx64 dimension, which contains multiple hidden layers, each layer contains multiple neurons, each neuron receives the input of the current moment and several previous moments, after the linear transformation of weights and biases, it is nonlinearly transformed through the activation function, and finally uses the adam optimizer and back propagation algorithm to output two types of discrimination results after 100 rounds of training optimization, namely short-circuit discharge and normal.
[0299] 5. Abnormal short circuit discharge sound detection of transmission lines
[0300] Figure 5 The flow chart of abnormal short-circuit discharge sound detection for power transmission lines provided in the embodiments of this specification, the device continuously collects sound information around the transmission line, and the collected sound is first subjected to a series of preprocessing; then the Mel spectrum features and Mel cepstral coefficient features are extracted respectively; next, the above-mentioned fusion voiceprint feature extraction method is used to obtain the voiceprint features that are finally input into the time-delay neural network model; the final model uses the fusion voiceprint features for prediction to determine whether the current sound is generated under short-circuit discharge conditions. When it is determined to be a short-circuit discharge sound, an alarm is issued, otherwise no measures are taken.
[0301] Figure 6 A schematic diagram of a transmission line short circuit identification device provided for one or more embodiments of this specification includes:
[0302] at least one processor; and,
[0303] a memory communicatively connected to the at least one processor; wherein,
[0304] The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:
[0305] Collect current sound information within the preset range of the transmission line;
[0306] Extracting current Mel-spectrogram features and current Mel-cepstral coefficient features of the current sound information;
[0307] Constructing a current fused voiceprint feature according to the current Mel-frequency spectrum feature and the current Mel-frequency cepstral coefficient feature;
[0308] The current fused voiceprint feature is input into a pre-trained short-circuit recognition neural network model, and a recognition result is output, so as to determine whether the current sound information is generated by short-circuit discharge according to the recognition result.
[0309] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.
Claims
1. A method for identifying a short circuit in a transmission line, characterized in that: include: Collect current sound information within the preset range of the transmission line; Extracting current Mel-spectrogram features and current Mel-cepstral coefficient features of the current sound information; Constructing a current fused voiceprint feature according to the current Mel-frequency spectrum feature and the current Mel-frequency cepstral coefficient feature; The current fused voiceprint feature is input into a pre-trained short-circuit recognition neural network model, and a recognition result is output, so as to determine whether the current sound information is generated by short-circuit discharge according to the recognition result.
2. The method according to claim 1, characterized in that Before inputting the current fused voiceprint feature into a pre-trained short-circuit recognition neural network model, the method further includes: Collecting transmission line short-circuit discharge sound data; Extracting Mel-spectrogram features and Mel-cepstral coefficient features of the transmission line short-circuit discharge sound data; Constructing a fused voiceprint feature of the power transmission line short-circuit discharge sound data according to the Mel-frequency spectrum feature and the Mel-frequency cepstrum coefficient feature; According to the fused voiceprint features of the transmission line short-circuit discharge sound data, a pre-created initial short-circuit recognition neural network model is trained to obtain a short-circuit recognition neural network model that meets the requirements.
3. The method according to claim 2, characterized in that The transmission line short-circuit discharge sound data includes: real transmission line short-circuit discharge sound data, simulated short-circuit discharge sound data and expanded short-circuit discharge sound data; The real power transmission line short-circuit discharge sound data includes multiple types of real power transmission line short-circuit discharge sound data; The simulated short-circuit discharge sound data is generated based on a mixture of different types of real power transmission line short-circuit discharge sound data; The expanded short-circuit discharge sound data is generated based on different types of real power transmission line short-circuit discharge sound data and the simulated short-circuit discharge sound data, and is aliased with noise in a power transmission line scenario.
4. The method according to claim 2, characterized in that: The step of extracting Mel-spectrogram features and Mel-cepstral coefficient features of the transmission line short-circuit discharge sound data includes: According to a preset pre-emphasis system, pre-emphasis processing is performed on the power transmission line short-circuit discharge sound data to obtain pre-emphasized power transmission line short-circuit discharge sound data; Dividing the pre-emphasized power transmission line short-circuit discharge sound data into multiple frames of sound data according to time; Multiply each frame of sound data by a preset window function to obtain windowed transmission line short-circuit discharge sound data; The Mel-spectrogram features and Mel-cepstral coefficient features of the windowed transmission line short-circuit discharge sound data are extracted.
5. The method according to claim 2, characterized in that: The step of extracting Mel-spectrogram features and Mel-cepstral coefficient features of the transmission line short-circuit discharge sound data includes: Performing Fourier transformation on each frame of the transmission line short-circuit discharge sound data to obtain Fourier transformation data, and obtaining a power spectrum signal based on the Fourier transformation data; Each filter in the Mel filter bank is point-multiplied with the power spectrum signal to obtain the Mel spectrum feature; The logarithm of the power spectrum signal after the Mel filter bank is taken to obtain a logarithmic processing result, and the logarithmic processing result is subjected to discrete cosine transformation to obtain a Mel cepstral coefficient feature.
6. The method according to claim 3, characterized in that The step of constructing a fused voiceprint feature of the power transmission line short-circuit discharge sound data according to the Mel frequency spectrum feature and the Mel cepstral coefficient feature includes: Selecting a real transmission line short-circuit discharge sound data from the transmission line short-circuit discharge sound data as reference sound data, and extracting a reference Mel-frequency spectrum feature and a reference Mel-frequency cepstral coefficient feature of the reference sound data; Extracting Mel-spectrogram features and Mel-cepstral coefficient features of the same feature dimension from all short-circuit discharge sounds in the transmission line short-circuit discharge sound data; Calculate feature similarities of the Mel-spectrogram features and Mel-cepstral coefficient features of all short-circuit discharge sounds and the benchmark Mel-spectrogram features and benchmark Mel-cepstral coefficient features of the benchmark sound data, respectively, to obtain an average feature similarity score of the Mel-spectrogram features and an average feature similarity score of the Mel-cepstral coefficient features; Constructing a feature fusion weight corresponding to the Mel-spectrogram feature and a feature fusion weight corresponding to the Mel-cepstral coefficient feature required for fusion of voiceprint features according to the average feature similarity score of the Mel-spectrogram feature and the average feature similarity score of the Mel-cepstral coefficient feature; A fused voiceprint feature of the power transmission line short-circuit discharge sound data is constructed based on the Mel-spectrogram feature, the feature fusion weight corresponding to the Mel-spectrogram feature, the Mel-cepstral coefficient feature, and the feature fusion weight corresponding to the Mel-cepstral coefficient feature.
7. The method according to claim 6, characterized in that The constructing the feature fusion weights corresponding to the Mel-spectrogram features and the feature fusion weights corresponding to the Mel-spectrogram coefficient features required for fusion voiceprint features according to the feature similarity average scores of the Mel-spectrogram features and the feature similarity average scores of the Mel-cepstral coefficient features, comprises: According to the formula and Construct the feature fusion weights corresponding to the Mel spectrum graph features and the feature fusion weights corresponding to the Mel cepstral coefficient features required for fusion voiceprint features; Among them, α is the feature fusion weight corresponding to the Mel-spectrogram feature, β is the feature fusion weight corresponding to the Mel-cepstral coefficient feature, avg1 is the average feature similarity score of the Mel-spectrogram feature, and avg2 is the average feature similarity score of the Mel-cepstral coefficient feature.
8. The method according to claim 6, characterized in that The constructing the current fused voiceprint feature according to the current Mel-frequency spectrum feature and the current Mel-frequency cepstral coefficient feature includes: The current fused voiceprint feature is constructed according to the current Mel-spectrogram feature, the feature fusion weight corresponding to the Mel-spectrogram feature, the current Mel-cepstral coefficient feature, and the feature fusion weight corresponding to the Mel-cepstral coefficient feature.
9. The method according to claim 6, characterized in that The step of constructing a fused voiceprint feature of the power transmission line short-circuit discharge sound data according to the Mel-spectrogram feature, the feature fusion weight corresponding to the Mel-spectrogram feature, the Mel-cepstral coefficient feature, and the feature fusion weight corresponding to the Mel-cepstral coefficient feature includes: According to the formula fmix=α*f1+β*f2, a fusion voiceprint feature of the transmission line short-circuit discharge sound data is constructed; Among them, fmix is the fused voiceprint feature of the transmission line short-circuit discharge sound data, α is the feature fusion weight corresponding to the Mel spectrum feature, β is the feature fusion weight corresponding to the Mel cepstral coefficient feature, f1 is the Mel spectrum feature, and f2 is the Mel cepstral coefficient feature.
10. A power transmission line short circuit identification device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: Collect current sound information within the preset range of the transmission line; Extracting current Mel-spectrogram features and current Mel-cepstral coefficient features of the current sound information; Constructing a current fused voiceprint feature according to the current Mel-frequency spectrum feature and the current Mel-frequency cepstral coefficient feature; The current fused voiceprint feature is input into a pre-trained short-circuit recognition neural network model, and a recognition result is output, so as to determine whether the current sound information is generated by short-circuit discharge according to the recognition result.
